fix(cambricon): prevent duplicate half definitions

src/CMakeLists.txt

@@ -424,6 +424,10 @@ if(WITH_CAMBRICON)
         target_compile_options(infiniops PUBLIC 
             "$<$<COMPILE_LANGUAGE:CXX>:SHELL:-idirafter /usr/local/neuware/lib/clang/11.1.0/include>"
         )
+        set(_cambricon_host_bang_compat_option
+            "$<$<COMPILE_LANGUAGE:CXX>:SHELL:-include ${CMAKE_CURRENT_SOURCE_DIR}/native/cambricon/host_bang_compat.h>")
+        target_compile_options(infiniops PRIVATE
+            "${_cambricon_host_bang_compat_option}")
     endif()
 
     list(APPEND DEVICE_LIST "cambricon")
@@ -1507,6 +1511,11 @@ if(GENERATE_PYTHON_BINDINGS)
         "$<$<COMPILE_LANGUAGE:CXX,CUDA>:-Wno-deprecated-declarations>"
     )
 
+    if(WITH_CAMBRICON AND CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+        target_compile_options(ops PRIVATE
+            "${_cambricon_host_bang_compat_option}")
+    endif()
+
     if(GENERATE_OPERATOR_CALL_INSTANTIATIONS)
         target_compile_definitions(ops PRIVATE
             INFINI_OPS_USE_OPERATOR_CALL_INSTANTIATIONS=1)

src/native/cambricon/host_bang_compat.h

@@ -0,0 +1,34 @@
+#ifndef INFINI_OPS_NATIVE_CAMBRICON_HOST_BANG_COMPAT_H_
+#define INFINI_OPS_NATIVE_CAMBRICON_HOST_BANG_COMPAT_H_
+
+// Neuware 6.0 leaves __BANG_HOSTDEVICE__ empty when its FP16/BF16 headers are
+// parsed by a host C++ compiler. Consequently, the function definitions in
+// those headers have external linkage and collide across translation units.
+// Pre-include each header with the BANG annotations mapped to standard C++
+// inline semantics. The vendor include guards then keep later cnrt/cnnl
+// includes from emitting the non-inline definitions.
+#if !defined(__BANGCC__) && !defined(__CUDACC__)
+#pragma push_macro("__BANGCC__")
+#pragma push_macro("__mlu_host__")
+#pragma push_macro("__mlu_func__")
+#define __BANGCC__ 1
+#define __mlu_host__
+#define __mlu_func__
+#include <bang_fp16.h>
+#pragma pop_macro("__mlu_func__")
+#pragma pop_macro("__mlu_host__")
+#pragma pop_macro("__BANGCC__")
+
+#pragma push_macro("__BANGCC__")
+#pragma push_macro("__mlu_host__")
+#pragma push_macro("__mlu_func__")
+#define __BANGCC__ 1
+#define __mlu_host__ inline
+#define __mlu_func__
+#include <bang_bf16.h>
+#pragma pop_macro("__mlu_func__")
+#pragma pop_macro("__mlu_host__")
+#pragma pop_macro("__BANGCC__")
+#endif
+
+#endif

tests/test_generate_wrappers.py

@@ -1133,3 +1133,19 @@ def test_shared_triton_config_parser_has_one_explicit_schema():
 
     assert "namespace infini::ops::triton::jit" in parser_header
     assert "namespace detail" in parser_header
+
+
+def test_cmake_configures_cambricon_half_compatibility():
+    root = pathlib.Path(__file__).parents[1]
+    source_cmake = (root / "src" / "CMakeLists.txt").read_text(encoding="utf-8")
+    compatibility_header = (
+        root / "src" / "native" / "cambricon" / "host_bang_compat.h"
+    ).read_text(encoding="utf-8")
+
+    assert "native/cambricon/host_bang_compat.h" in source_cmake
+    assert "target_compile_options(infiniops PRIVATE" in source_cmake
+    assert "target_compile_options(ops PRIVATE" in source_cmake
+    assert source_cmake.count("_cambricon_host_bang_compat_option}") == 2
+    assert "#include <bang_fp16.h>" in compatibility_header
+    assert "#include <bang_bf16.h>" in compatibility_header
+    assert "#define __mlu_host__ inline" in compatibility_header